Python can now use the C99 NAN constant or __builtin_nan()

[vstinner]

BPO	46640
Nosy	@tim-one, @rhettinger, @mdickinson, @vstinner, @bitdancer, @encukou
PRs	bpo-46640: Py_NAN now uses the C99 NAN constant #31134
Files	test_nan_bits.py

^{Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.}

Show more details

GitHub fields:

assignee = None
closed_at = <Date 2022-02-25.15:09:43.947>
created_at = <Date 2022-02-04.20:24:09.387>
labels = ['interpreter-core', '3.11']
title = 'Python can now use the C99 NAN constant or __builtin_nan()'
updated_at = <Date 2022-02-25.15:09:43.946>
user = 'https://github.com/vstinner'

bugs.python.org fields:

activity = <Date 2022-02-25.15:09:43.946>
actor = 'vstinner'
assignee = 'none'
closed = True
closed_date = <Date 2022-02-25.15:09:43.947>
closer = 'vstinner'
components = ['Interpreter Core']
creation = <Date 2022-02-04.20:24:09.387>
creator = 'vstinner'
dependencies = []
files = ['50603']
hgrepos = []
issue_num = 46640
keywords = ['patch']
message_count = 17.0
messages = ['412531', '412533', '412536', '412538', '412539', '412540', '412542', '412576', '412581', '412582', '412624', '412626', '412728', '412743', '412747', '412748', '414002']
nosy_count = 6.0
nosy_names = ['tim.peters', 'rhettinger', 'mark.dickinson', 'vstinner', 'r.david.murray', 'petr.viktorin']
pr_nums = ['31134']
priority = 'normal'
resolution = 'fixed'
stage = 'resolved'
status = 'closed'
superseder = None
type = None
url = 'https://bugs.python.org/issue46640'
versions = ['Python 3.11']

[vstinner]

While debugging a GCC regression (*) on "HUGE_VAL * 0" used by Py_NAN macro, I noticed that Python can now C99 "NAN" constant.

(*) https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104389

In bpo-45440, I already removed legacy code for pre-C99 support and old platforms:

"Building Python now requires a C99 <math.h> header file providing the following functions: copysign(), hypot(), isfinite(), isinf(), isnan(), round()."

Attached patch modifies Py_NAN to simply reuse NAN.

mathmodule.c and cmathmodule.c m_nan() still use _Py_dg_stdnan() by default (if PY_NO_SHORT_FLOAT_REPR is not defined).

[vstinner]

The Py_NAN has a special implementation for the ICC compiler, if __INTEL_COMPILER and ICC_NAN_STRICT macros are defined, in bpo-21167:
---
commit edbc28c
Author: R David Murray <rdmurray@bitdance.com>
Date: Thu Aug 13 09:58:07 2015 -0400

bpo-21167: Fix definition of NAN when ICC used without -fp-model strict.

Patch from Chris Hogan of Intel, reviewed by Mark Dickinson.

---

I don't know if it should be kept if Py_NAN is modified to use the NAN constant. In case of doubt, I prefer to remove the ICC code since we have no ICC buildbot anymore and maybe ICC changed in the meanwhile.

Also, I don't have acess to ICC.

[vstinner]

mathmodule.c and cmathmodule.c m_nan() still use _Py_dg_stdnan() by default (if PY_NO_SHORT_FLOAT_REPR is not defined).

These functions are only use to create the following constants:

• math.nan
• cmath.nan
• cmath.nanj

[vstinner]

Manual test to check if m_nan(), _Py_dg_stdnan(0) and Py_NAN are exactly the same number (same bits):

$ ./python
>>> import math, struct
>>> m_nan=math.nan; Py_NAN=math.atan2(m_nan, 1.0)
>>> Py_NAN is m_nan
False
>>> struct.pack('d', m_nan) == struct.pack('d', Py_NAN)
True
>>> struct.pack('d', Py_NAN)
b'\x00\x00\x00\x00\x00\x00\xf8\x7f'

=> see attached script: test_nan_bits.py

"struct.pack('d', m_nan) == struct.pack('d', Py_NAN)" is true with #75317 on Fedora 35 with gcc-11.2.1-7.fc35.x86_64. I tested with "gcc -O0" and "gcc -O3".

GCC float.h defines NAN with:

    #define NAN (__builtin_nanf (""))

GCC: "Built-in Function: double __builtin_nan (const char *str): This is an implementation of the ISO C99 function nan."

[vstinner]

Python/dtoa.c uses:

/* Standard NaN used by _Py_dg_stdnan. */

#define NAN_WORD0 0x7ff80000
#define NAN_WORD1 0

/* Return a 'standard' NaN value.

There are exactly two quiet NaNs that don't arise by 'quieting' signaling
NaNs (see IEEE 754-2008, section 6.2.1). If sign == 0, return the one whose
sign bit is cleared. Otherwise, return the one whose sign bit is set.
*/

double
_Py_dg_stdnan(int sign)
{
    U rv;
    word0(&rv) = NAN_WORD0;
    word1(&rv) = NAN_WORD1;
    if (sign)
        word0(&rv) |= Sign_bit;
    return dval(&rv);
}

[vstinner]

see attached script: test_nan_bits.py

test_nan_bits.py says "True" for Python built with "clang -O3" and with "clang -O0".

[vstinner]

Using clang -E, I see that clang also replaces NAN with: __builtin_nanf ("").

[mdickinson]

The big blocker here is that a platform that fully supports C99 might not define the "NAN" macro. I don't think we can require that NAN be defined in order for Python to build (which is what the PR currently does, if I'm understanding it correctly).

Python deliberately doesn't assume IEEE 754 floating-point. By requiring that the C "NAN" macro is present to be able to build Python, we'd be effectively requiring IEEE 754 by stealth. (No other common floating-point format has NaNs.)

I'd be fully on board with a decision to require IEEE 754 floating-point for Python in future, but that decision would at least need a python-dev discussion - we shouldn't sneak it in by the back door.

[vstinner]

The big blocker here is that a platform that fully supports C99 might not define the "NAN" macro. I don't think we can require that NAN be defined in order for Python to build (which is what the PR currently does, if I'm understanding it correctly).

If a platform doesn't implement NaN, it should define the Py_NO_NAN macro:

/* Py_NAN
 * A value that evaluates to a NaN. On IEEE 754 platforms INF*0 or
 * INF/INF works. Define Py_NO_NAN in pyconfig.h if your platform
 * doesn't support NaNs.
 */
#if !defined(Py_NAN) && !defined(Py_NO_NAN)
   // Use C99 "NAN" constant: quiet Not-A-Number (when supported)
#  define Py_NAN NAN
#endif

Or do you mean that a platform can support NaN but don't define the <math.h> NAN macro?

[mdickinson]

If a platform doesn't implement NaN, it should define the Py_NO_NAN macro

Ah. In that case your PR description (and the PR news entry) is misleading:

Building Python now requires a C99 <math.h> header file providing the
NAN constant.

Please could you update them?

[vstinner]

New changeset 54842e4 by Victor Stinner in branch 'main':
bpo-46640: Py_NAN now uses the C99 NAN constant (GH-31134)
54842e4